User personalized content pushing system and promotion method based on big data

Abstract

The invention relates to the technical field of user personalized content pushing, in particular to a user personalized content pushing system based on big data and a promotion method. The method comprises the following steps: collecting user behavior and preference data through multiple channels; dynamically constructing user portraits, and enriching the user portraits in combination with multi-dimensional data; based on user portraits and content features, a recommendation model is constructed by using a deep neural network, and user preferences are accurately predicted through multiple hidden layers; through a push model and a multi-objective optimization method, the probability of pushed content is calculated in combination with indexes such as a click rate and a conversion rate, the push frequency and content selection are optimized, and the correlation of the pushed content and the user satisfaction are ensured; by monitoring user feedback in real time and adjusting a pushing strategy, intelligent decision and personalized recommendation of content pushing are realized, and long-term accurate pushing and improvement of user experience are ensured. According to the method, the accurate user portrait is constructed through multi-data source collection and processing, so that personalized content pushing is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of user personalized content push, and in particular to a user personalized content push system and promotion method based on big data. Background Art

[0002] With the widespread application of the Internet and smart devices, the amount of data generated by users is growing exponentially. How to efficiently extract valuable information from massive data and make personalized content recommendations based on user needs has become an important issue facing major Internet companies. Traditional recommendation systems usually rely on static user portraits and rule-driven recommendation algorithms, which are difficult to adapt to changes in user interests in real time, resulting in unsatisfactory recommendation results.

[0003] A Chinese invention patent with announcement number CN116522011B discloses a push method and push system based on big data. The method establishes a two-layer push knowledge base, including text features and boundary contour features, and extracts keywords from the target user's historical behavior data and extracts boundaries from the input training image data to reversely restore the corresponding search terms, and retrieves the corresponding content from the Internet to push information. The technical solution provided by the patent has the advantages of high efficiency, accuracy, expansion and personalization, is suitable for various information push scenarios, and has practicality and promotion value.

[0004] In order to improve the accuracy of recommendations and user satisfaction, personalized content recommendation methods based on big data have emerged. They aim to build dynamic user portraits in real time by deeply mining user behavior data, and combine advanced machine learning and deep learning algorithms to achieve more intelligent recommendation decisions. In particular, with the rapid development of artificial intelligence and deep neural network technologies, recommendation systems can not only process large-scale data, but also accurately capture changes in user interests through multi-level feature fusion and predictive modeling. At the same time, the optimization and frequency regulation of pushed content have also become key factors in improving user experience. Existing push strategies mostly remain at simple frequency control and content selection, and have not yet fully utilized real-time feedback mechanisms and multi-objective optimization strategies. Summary of the invention

[0005] The purpose of the present invention is to address the problems existing in the background technology and to propose a user personalized content push system and promotion method based on big data.

[0006] The technical solution of the present invention is a method for promoting user personalized content based on big data, comprising the following specific implementation steps: S1. Collect user behavior and preference data, use automated data collection tools to obtain data, and use unified data format conversion tools to format heterogeneous data, converting the collected user behavior and preference data into a processable standardized data structure; S2. Perform quality checks on user behavior and preference data, remove duplicate, erroneous or incomplete data, fill in missing data, remove abnormal data, standardize the processed data, and adjust the data to a uniform scale; S3. Dynamic personalized content recommendation method based on user behavior analysis, capturing user interest changes, defining and normalizing behavior weights, combining time decay factors to form user interest vectors, using K-Means clustering algorithm to classify interest vectors, and generating priority rankings based on user interest tags, dynamically adjusting the profile according to user behavior changes, updating the interest vector, constructing the final user profile, and generating verification marks for the final user profile; S4. Verify the rationality of the received user portrait, build a training data set, use a negative sampling strategy to balance the data set, generate user-content interaction samples, and then fuse the user portrait and content features to build a recommendation model that fuses the user portrait and content features, extract high-dimensional features and predict user preferences for content, and generate personalized recommended content; S5. Calculate the push probability of each recommended content, and optimize the push content selection based on the user's click-through rate, conversion rate, and interaction rate. Generate a personalized push list by adjusting the push frequency and push it to the user. Meanwhile, monitor the user's feedback on the push content in real time and adjust the content and push frequency.

[0007] Preferably, the process of eliminating duplicate, erroneous or incomplete data is as follows: S21. For duplicate data, use the hash value matching method: h(x)=H0(x); Where h(x) represents the hash value of data x; H0(x) represents the standard hash function; Then store the hash value in the data list List, check whether the same value already exists in the data list List, and if so, mark it as duplicate data; S22. For data fields with a high proportion of missing values, calculate the field missing rate: ; In the formula, R miss N represents the missing rate; miss Indicates the number of missing values; N total Indicates the total number of data items; If R miss >θ1, the field is marked as unavailable; Among them, θ1 represents the set missing rate threshold.

[0008] Preferably, the process of generating priority ranking based on user interest tags is as follows: S31. For each type of behavior b of any user u, define its behavior weight w(b): ; In the formula, f(b) represents the frequency of the user performing behavior b; v(b) represents the importance weight of behavior b; T u Represents the total frequency of all behaviors of user u; S32. Normalize behavior weights to eliminate dimensional effects: ; Where w'(b) represents the normalized behavior weight; max(w) and min(w) represent the maximum and minimum values ​​of the behavior weight, respectively; S33. Based on the time effect of behavior, recent behavior is given a higher weight and the time decay weight d(b,t) is defined as: ; In the formula, λ represents the time decay coefficient, which reflects the speed at which the importance of behavior decays over time; Δt represents the time difference between behavior b and the current time; S34. Combining user behavior weight and time decay, construct interest vector I u : ; Where w'(b) represents the normalized behavior b i The behavior weight of ; d(b,t) represents the behavior b i The time decay weight of L i Indicates behavior i The corresponding interest category label; S35. Use K-Means clustering to divide user interest vectors into different categories: ; In the formula, C k represents the kth interest category group; μ k represents the centroid of the k-th interest category group category; S36. Weight the user's interest tags and generate a priority ranking: ; In the formula, S u It represents the sorted interest tag set, which is arranged from high to low by weight; Sort() represents the sorting algorithm; S37. Based on the priority-sorted user interest tags, dynamically adjust the profile according to changes in user behavior.

[0009] Preferably, the adjustment process of dynamically adjusting the portrait is as follows: S41. For behavior b, adjust the weight based on historical data and real-time behavior data: ; In the formula, w t (b) represents the current weight; w t-1 (b) represents the historical weight; w new (b) represents the weight of the latest behavior calculation; γ represents the weight coefficient of historical data, ranging from [0,1]; S42, Dynamic Synthesis Interest Vector: ; In the formula, represents the updated interest vector; I u Represents the original interest vector; I new represents the interest vector generated by the newly added behavior; α and β represent dynamic weighting coefficients, satisfying α+β=1; S43. Construct a unified feature vector to generate a final user portrait.

[0010] Preferably, the process of generating the final user portrait is as follows: S51. Constructing a multi-dimensional user feature vector: ; In the formula, f1, f2, ..., f k Indicates demographic characteristics; S52. Add image, text, and voice feature embedding to enrich user portraits: ; In the formula, G u represents the image feature embedding vector; T u Represents the text feature embedding vector; A u Represents the speech feature embedding vector; S53. Reduce dimension through principal component analysis and improve computational efficiency: ; In the formula, represents the user feature vector after dimensionality reduction; W represents the dimensionality reduction mapping matrix; S54. Output end-user portrait .

[0011] Preferably, the generation process of the verification mark is as follows: S61. Randomly select large prime numbers p and q where p ≠ q, and calculate the first-level marking generation factor fg1: fg1 = p × q × (p - 1) × (q - 1); where p > 3 and q > 3; S62. Randomly select an integer r and calculate the second-level marking generation factor fg2: ; where p > 2 r and q > 2 r ; S63. Select the marking parsing code Ca and the marking generation code Cg such that: Cg × Ca = 1 (mod n); where 1 < Ca < fg2, gcd(Ca, fg2) = 1, and gcd() represents the greatest common divisor function; S64. Generate the verification mark La = H1(M) Cg mod fg1; where M represents the binary string of the constructed user profile ; H1() represents the hash function.

[0012] Preferably, the verification process for verifying the rationality of the received user profile is as follows: S71. Calculate the first-level verification factor fa1 = (La) Ca mod fg1; where Ca represents the marking parsing code; fg1 represents the first-level marking generation factor; S72. Calculate the second-level verification factor fa2 = H1(M); where M represents the binary string of the constructed user profile ; H1() represents the hash function; S73. If fa1 = fa2, it indicates that the received user profile is reasonable.

[0013] Preferably, the network structure of the recommendation model for fusing the user profile and content features includes: S81. The recommendation model for fusing the user profile and content features is based on a deep neural network architecture, including: an input layer, a hidden layer, and an output layer: ; In the formula, represents the preference probability of user u for content i predicted by the model; σ represents the activation function; W (L) and b (L) respectively represent the weight matrix and bias vector of the L-th layer; h (L-1) represents the output of the (L - 1)-th layer; S82, feature fusion layer, user features and content features interact in the feature fusion layer: ; Where W u and W c The projection matrix representing user features and content features; represents a nonlinear activation function; b (1) represents the bias term of the feature fusion layer; h (1) represents the output of the first layer, i.e., the fusion result of user and content features; S83, hidden layer extracts high-dimensional features: ; Where l represents the number of the hidden layer, l=2, 3, ..., L-1; S84. Convert the recommendation model output into preference probability through the Sigmoid function: ; In the formula, It represents the probability of user u's preference for content i predicted by the model.

[0014] Preferably, the process of generating a personalized push list is as follows: S91. For each recommended content C i , based on user portraits and the predicted value of the recommended model , calculate the probability P of push push (C i ||u): ; In the formula, λ'' represents the attenuation coefficient; S92, combining the user's click rate, conversion rate and user interaction rate, and combining the objective function L push Select push content: ; Where, L click Indicates click rate loss; L conversion Indicates conversion rate loss; L engagement represents the user interaction rate loss; α', β' and γ' represent the weight coefficients of each objective function respectively; S93. For each user u, calculate a reasonable push frequency f u : ; Where σ' represents the adjustment parameter; C u represents the set of recommended content received by user u; S94: Push probability P based on each contentpush (C i ||u), user interest prediction value and the properties of the content I(C i ), comprehensively generate personalized push list List u : ; In the formula, I(C i ) represents a content attribute vector, including but not limited to the freshness, category, and tag information of the content; Sort() represents a sorting function.

[0015] The technical solution of the present invention is a user personalized content push system based on big data, which is used to execute the above-mentioned user personalized content promotion method based on big data, including: The data collection module is responsible for collecting user behavior and preference data and transmitting it to the data processing module for processing; Data processing module, used to clean the collected user behavior and preference data; The data analysis module is used to accurately model and build multi-dimensional user portraits based on user behavior and preference data using cluster analysis and association rule analysis techniques; Personalized recommendation module, which is used to build a recommendation model and generate personalized recommendation content based on user portraits, content features and real-time user behavior; The content push module is used to push personalized content to users through an intelligent push mechanism, and can automatically optimize and adjust during the push process.

[0016] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects: The present invention designs a user personalized content push system and promotion method based on big data. By innovatively integrating data collection, user portraits, deep learning and real-time feedback mechanisms, an efficient, intelligent and accurate personalized content push system is realized, which greatly improves the user experience, recommendation effect and system optimization ability, and has significant technical advantages: (1) Improve data processing efficiency and accuracy: Through the automated collection and unified formatting of multi-source data, the high quality and consistency of data are ensured, and data redundancy, errors and missing issues are effectively avoided, thus providing an accurate data foundation for subsequent analysis and modeling; (2) Accurately construct user portraits: By dynamically adjusting user behavior weights, time effects, and comprehensive analysis of multi-dimensional features, accurate modeling of user interests is achieved, which can reflect changes in user interests in real time. This approach ensures the real-time and accuracy of personalized recommendations and enhances user experience. (3) Optimizing the performance of the recommendation model: The deep neural network (DNN) model is used to learn the user portrait and content features, and the L2 regularization, Dropout and other anti-overfitting technologies are used to improve the generalization ability of the model and avoid the decline in recommendation effect caused by overfitting. In addition, the dynamic learning rate adjustment and optimization algorithm improve the training efficiency and stability of the model; (4) Improving the relevance of recommended content and user satisfaction: Through a multi-objective optimization model, multi-dimensional indicators such as click-through rate, conversion rate, and interaction rate are comprehensively considered to achieve accurate selection of push content, avoid excessive push, and at the same time improve the relevance of recommended content and user satisfaction; (5) Continuously optimize push strategy: Real-time monitoring of user feedback on recommended content, and adjustment of push content and frequency based on feedback, can dynamically adjust the recommendation strategy, thereby enhancing the adaptability and intelligence of the system and ensuring the long-term effectiveness of the recommendation; (6) Ensure the security and credibility of user portraits and data integrity: Check marks are used to verify the validity of user portraits and prevent tampering or data loss during data transmission. By generating and checking marks, it can be ensured that the transmitted user portraits are still legal and have not been tampered with after passing through the network or other systems; check marks are generated through hash functions and mathematical operations, so that the integrity of user portraits is guaranteed. Any tampering with the user portrait will cause the check mark to fail to match, thereby generating a warning in the system to prevent erroneous or illegal data from entering the subsequent processing links. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a system architecture diagram of a user personalized content push system based on big data proposed by the present invention; Figure 2 This is a method flow chart of a method for promoting user personalized content based on big data proposed by the present invention. DETAILED DESCRIPTION

[0018] Embodiment 1, as Figure 1 As shown, the present invention proposes a user personalized content push system based on big data, including: a data collection module, a data processing module, a data analysis module, a personalized recommendation module and a content push module.

[0019] The data collection module is responsible for collecting user behavior and preference data, including but not limited to behavior data, preference data, historical interaction data and social network information, and transmitting it to the data processing module for processing; The data processing module cleans the collected user behavior and preference data; The data analysis module uses cluster analysis and association rule analysis techniques to accurately model and construct multi-dimensional user portraits based on user behavior and preference data; The personalized recommendation module builds a recommendation model based on user portraits, content features, and real-time user behavior to generate personalized recommendation content; The content push module pushes personalized content to users through an intelligent push mechanism. The push forms include but are not limited to notifications, emails, and message reminders, and can automatically optimize and adjust during the push process.

[0020] Embodiment 2, as Figure 2 As shown, a user personalized content promotion method based on big data proposed by the present invention is applied to a user personalized content push system based on big data proposed in Example 1, and its specific implementation steps are as follows: S1. Collect user behavior and preference data from multiple data sources (including but not limited to user operation logs, search engine behavior data, social media data, and device sensor data), and use automated data collection tools (including but not limited to Web crawlers, API interfaces, and sensor gateways) to obtain data regularly or in real time, and use unified data format conversion tools to format heterogeneous data sources and convert the collected user behavior and preference data into a processable standardized data structure; The user behavior and preference data are then transmitted to the data processing module.

[0021] S2. The data processing module system performs quality checks on user behaviors and preferences, removes duplicate, erroneous or incomplete data, and processes missing data through missing value filling algorithms to improve data integrity. It also uses anomaly detection algorithms to remove abnormal data to ensure data reliability and stability. The specific implementation process is as follows: S21. Identify and mark duplicate, missing, abnormal or invalid data in the data set. The identification process is as follows: S2101. For duplicate data, use the hash value matching method: h(x)=H0(x); Where h(x) represents the hash value of data x; H0(x) represents the standard hash function; Then store the hash value in the data list List, check whether the same value already exists in the data list List, and if so, mark it as duplicate data; S2102. For data fields with a high proportion of missing values, calculate the field missing rate: ; In the formula, R miss N represents the missing rate; miss Indicates the number of missing values; N totalIndicates the total number of data items; If R miss >θ1, the field is marked as unavailable; Among them, θ1 represents the missing rate threshold set; S22. Identify missing value data types and fill in missing values ​​to improve data integrity: For numerical data, the mean filling method is used to replace the missing values ​​with the calculated mean: ; In the formula, x i Indicates missing values; X miss represents the set of missing values; N represents the number of non-missing data points; x j represents non-missing values; For time series data, linear interpolation is used: ; In the formula, x t Indicates missing point value; x t-1 and x t+1 Indicates the nearest known value before and after the missing point; t t-1 and t t+1 Indicates the corresponding timestamp; S23, detect and remove or replace outliers, the processing process is as follows: S2301, using Z-score based outlier detection: ; In the formula, Z i Represents data point x i The Z score; μ represents the data mean; σ represents the data standard deviation; S2302, if | Z i |>θ2, it is marked as an outlier; In the formula, θ2 represents the set Z score threshold; Treatment of replacing outliers: If the proportion of outliers is low, they are directly removed; if the proportion of outliers is high, the median is used instead, x new =Median(X); In the formula, X represents the data set; Median() means taking the median; S24. Adjust the data to a uniform scale to avoid the impact of dimensional differences on model performance. The data standardization formula is: ; In the formula, x min and x max Respectively represent the minimum and maximum values ​​of the data; represents the normalized value; S25, transmitting the processed data to the data analysis module.

[0022] S3, the data analysis module constructs user portraits in real time by dynamically adjusting behavior weights, clustering interest tags, and constructing feature vectors. The construction process is as follows: S31. Extract behavioral data from user data and extract core features based on user behavior records: S3101. For each type of behavior b (including but not limited to click, favorite, purchase) of any user u, define its behavior weight w(b): ; In the formula, f(b) represents the frequency of the user performing behavior b; v(b) represents the importance weight of behavior b; T u Represents the total frequency of all behaviors of user u; S3102. Normalize behavior weights to eliminate dimensional effects: ; Where w'(b) represents the normalized behavior weight; max(w) and min(w) represent the maximum and minimum values ​​of the behavior weight, respectively; S3103. Consider the time effect of behavior, give higher weight to recent behavior, and define the time decay weight d(b,t): ; In the formula, λ represents the time decay coefficient, which reflects the speed at which the importance of behavior decays over time; Δt represents the time difference between behavior b and the current time; S32, mapping user interest tags through behavioral characteristics, and performing accurate interest classification. The classification process is as follows: S3201, integrating user behavior weight and time decay, constructing interest vector I u : ; Where w'(b) represents the normalized behavior b i The behavior weight of ; d(b,t) represents the behavior b i The time decay weight of L i Indicates behavior i The corresponding interest category label; S3202, using K-Means clustering to divide user interest vectors into different categories: ; In the formula, C k represents the kth interest category group; μ k Represents the centroid of the category, updated dynamically; S3203. Weight the user's interest tags and generate a priority ranking: ; In the formula, S u It represents the sorted interest tag set, which is arranged from high to low by weight; Sort() represents the sorting algorithm; S33. Dynamically adjust the profile according to changes in user behavior. The adjustment process is as follows: S3301. For behavior b, adjust the weight based on historical data and real-time behavior data: ; In the formula, w t (b) represents the current weight; w t-1 (b) represents the historical weight; w new (b) represents the weight of the latest behavior calculation; γ represents the weight coefficient of historical data, ranging from [0,1]; S3302, Dynamic Synthesis Interest Vector: ; In the formula, represents the updated interest vector; I u Represents the original interest vector; I new represents the interest vector generated by the newly added behavior; α and β represent dynamic weighting coefficients, satisfying α+β=1; S34. Construct a unified feature vector. The construction process is as follows: S3401. Construct a multi-dimensional user feature vector: ; In the formula, f1, f2, ..., f k Indicates demographic characteristics (including but not limited to age, gender, region); S3402, add image, text, and voice feature embedding to enrich user portraits: ; In the formula, G u represents the image feature embedding vector; T u Represents the text feature embedding vector; A u Represents the speech feature embedding vector; S3403, reduce dimension through principal component analysis (PCA) and improve computational efficiency: ; In the formula, represents the user feature vector after dimensionality reduction; W represents the dimensionality reduction mapping matrix; S35. Build user portrait , and generate a check mark for it. The generation process is as follows: S3501. Randomly select large prime numbers p and q where p ≠ q, and calculate the first-level marking generation factor fg1: fg1 = p × q × (p - 1) × (q - 1); where p > 3 and q > 3; S3502. Randomly select an integer r and calculate the second-level marking generation factor fg2: ; where p > 2 r and q > 2 r ; S3503. Select the marking parsing code Ca and the marking generation code Cg such that: Cg × Ca = 1 (mod n); where 1 < Ca < fg2, gcd(Ca, fg2) = 1, and gcd() represents the greatest common divisor function; S3504. Generate the verification mark La = H1(M) Cg mod fg1; where M represents the binary string of the constructed user profile ; H1() represents the hash function; S36. Transmit {La, } to the personalized recommendation module.

[0023] S4. The personalized recommendation module constructs a training dataset through the fusion of the user profile and content features. The profile features of each user interact with the features of relevant content through the feature fusion layer to generate a preference prediction for the user and content. A recommendation model based on a deep neural network is designed to perform deep learning on user features and content features through multiple hidden layers, thereby predicting the preference probability of the user for the content. The specific implementation process is as follows: S41. Receive {La, }, and verify the rationality of the received user profile . The verification process is as follows: S4101. Calculate the first-level verification factor fa1 = (La) Ca mod fg1; where Ca represents the marking parsing code; fg1 represents the first-level marking generation factor; S4102. Calculate the second-level verification factor fa2 = H1(M); where M represents the binary string of the constructed user profile ; H1() represents the hash function; S4103. If fa1 = fa2, it means that the received user profile is reasonable; S42. Based on the user profile constructed in step S3 The training data set for the recommendation model is generated by combining historical interaction data. The training data set construction process is as follows: S4201, corresponding to each user u, combined with user portrait and content feature C i , generate training samples: ; In the formula, y ui represents the interaction label between user u and content i; if y ui =1, indicating interest; if y ui =0, indicating no interest; D indicates the training data set; S4202, to balance the data set (i.e., positive and negative samples are unbalanced), a negative sampling strategy is adopted: ; In the formula, f i represents the interaction frequency of content i. The higher the content, the lower the sampling frequency. p(sample) represents the probability of content i being sampled. S43. Build a recommendation model that integrates user portraits and content features, and train it with deep learning methods; (1) The recommendation model that integrates user profile and content features is based on a deep neural network (DNN) architecture, including input layer, hidden layer, and output layer: ; In the formula, represents the probability of user u's preference for content i predicted by the model; σ represents the activation function, and this embodiment adopts the Sigmoid activation function; W (L) and b (L) Represent the weight matrix and bias vector of the Lth layer respectively; h (L-1) represents the output of the L-1th layer; (2) Feature fusion layer: User features and content features interact at the feature fusion layer: ; Where W u and W c The projection matrix representing user features and content features; represents a nonlinear activation function, and this embodiment adopts the ReLU activation function; b (1) represents the bias term of the feature fusion layer; h (1) represents the output of the first layer, i.e., the fusion result of user and content features; (3) Multi-layer hidden layers extract high-dimensional features: ; Where l represents the number of the hidden layer, l=2, 3, ..., L-1; (4) Convert the model output into preference probability through the Sigmoid function: ; In the formula, represents the probability of user u's preference for content i predicted by the model; S44. Train the recommendation model based on user interaction data: S4401. Use the cross entropy loss function to measure the difference between the model prediction and the true interaction label: ; Where L represents the total loss; y ui represents the real interaction label; S4403, using adaptive gradient optimization algorithm to update model parameters: ; In the formula, θ represents the model parameter set; η represents the learning rate; Represents the gradient of the loss function with respect to the parameters; S45. To improve the performance and generalization ability of the model, the training process is optimized to avoid overfitting and improve computational efficiency: S4501. In order to prevent the model from overfitting, the L2 regularization term is introduced: ; Where λ' represents the regularization coefficient; ||θ|| 2 Represents the L2 norm of the model parameters; S4502. Add a Dropout layer to the hidden layer to randomly discard some neurons to prevent overfitting: ; In the formula, represents the output of the lth layer after Dropout; h (l) represents the original output of the lth layer; p' represents the discard probability; S46, dynamic learning rate adjustment formula: ; In the formula, represents the learning rate of the t'th iteration, which gradually decays as the training progresses; η0 represents the initial learning rate; k' represents the learning rate decay factor; t' represents the current number of iterations; S47: Call the recommendation model that integrates user portraits and content features to generate personalized recommendation content .

[0024] S5. The content push module realizes personalized content push through a comprehensive push probability model, multi-objective optimization and real-time feedback mechanism. The specific implementation process is as follows: S51. For each recommended content C i , based on user portraits and the predicted value of the recommended model , calculate the probability of push P push (C i ||u): ; In the formula, λ'' represents the attenuation coefficient; S52, combining the user's click rate, conversion rate and user interaction rate, and combining the objective function L push Select push content: ; Where, L click Indicates click rate loss; L conversion Indicates conversion rate loss; L engagement represents the user interaction rate loss; α', β' and γ' represent the weight coefficients of each objective function respectively; S53. For each user u, calculate a reasonable push frequency f u : ; Where σ' represents the adjustment parameter; C u represents the set of recommended content received by user u; S54: Push probability P based on each content push (C i ||u), user interest prediction value and the properties of the content I(C i ), comprehensively generate personalized push list List u : ; In the formula, I(C i ) represents a content attribute vector, including but not limited to the freshness, category, and tag information of the content; Sort() represents a sorting function; S55. Send personalized push list List u Push to user u; S56. Real-time monitoring of user u's feedback on the pushed content, and adjustment of the pushed content and frequency: ; In the formula, represents the predicted value of preference after feedback adjustment; Δy ui Indicates the user's content C i Deviation of feedback (including but not limited to clicks, favorites, and conversions); η feedback Represents the feedback adjustment coefficient, which controls the weight of real-time feedback.

[0025] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A user personalized content promotion method based on big data, characterized in that: The specific implementation steps include the following: S1. Collect user behavior and preference data, use automated data collection tools to obtain data, and use unified data format conversion tools to format heterogeneous data, converting the collected user behavior and preference data into a processable standardized data structure; S2. Perform quality checks on user behavior and preference data, remove duplicate, erroneous or incomplete data, fill in missing data, remove abnormal data, standardize the processed data, and adjust the data to a uniform scale; S3. Dynamic personalized content recommendation method based on user behavior analysis, capturing user interest changes, defining and normalizing behavior weights, combining time decay factors to form user interest vectors, using K-Means clustering algorithm to classify interest vectors, and generating priority rankings based on user interest tags, dynamically adjusting the profile according to user behavior changes, updating the interest vector, constructing the final user profile, and generating verification marks for the final user profile; S4. Verify the rationality of the received user portrait, build a training data set, use a negative sampling strategy to balance the data set, generate user-content interaction samples, and then fuse the user portrait and content features to build a recommendation model that fuses the user portrait and content features, extract high-dimensional features and predict user preferences for content, and generate personalized recommended content; S5. Calculate the push probability of each recommended content, and optimize the push content selection based on the user's click-through rate, conversion rate, and interaction rate. Generate a personalized push list by adjusting the push frequency and push it to the user. Meanwhile, monitor the user's feedback on the push content in real time and adjust the content and push frequency.

2. The method for promoting user personalized content based on big data according to claim 1, characterized in that: The process of eliminating duplicate, erroneous or incomplete data is as follows: S21. For duplicate data, use the hash value matching method: h(x)=H0(x); Where h(x) represents the hash value of data x; H0(x) represents the standard hash function; Then store the hash value in the data list List, check whether the same value already exists in the data list List, and if so, mark it as duplicate data; S22. For data fields with a high proportion of missing values, calculate the field missing rate: ； In the formula, R miss N represents the missing rate; miss Indicates the number of missing values; N total Indicates the total number of data items; If R miss >θ1, the field is marked as unavailable; Among them, θ1 represents the set missing rate threshold.

3. The method for promoting user personalized content based on big data according to claim 1, characterized in that: The process of generating priority ranking based on user interest tags is as follows: S31. For each type of behavior b of any user u, define its behavior weight w(b): ； In the formula, f(b) represents the frequency of the user performing behavior b; v(b) represents the importance weight of behavior b; T u Represents the total frequency of all behaviors of user u; S32. Normalize behavior weights to eliminate dimensional effects: ； In the formula, w'(b) represents the normalized behavior weight; max(w) and min(w) represent the maximum and minimum values ​​of the behavior weight, respectively; S33. Based on the time effect of behavior, recent behavior is given a higher weight and the time decay weight d(b,t) is defined as: ； In the formula, λ represents the time decay coefficient, which reflects the speed at which the importance of behavior decays over time; Δt represents the time difference between behavior b and the current time; S34. Combining user behavior weight and time decay, construct interest vector I u : ； Where w'(b) represents the normalized behavior b i The weight of the behavior; d(b,t) represents behavior b i The time decay weight of L i Indicates behavior i The corresponding interest category label; S35. Use K-Means clustering to divide user interest vectors into different categories: ； In the formula, C k represents the kth interest category group; μ k represents the centroid of the k-th interest category group category; S36. Weight the user's interest tags and generate a priority ranking: ； In the formula, S u It represents the sorted interest tag set, which is arranged from high to low by weight; Sort() represents the sorting algorithm; S37. Dynamically adjust the portrait according to the user interest tags sorted by priority and the changes in user behavior.

4. The method for promoting user personalized content based on big data according to claim 3, characterized in that: The adjustment process of dynamically adjusting the portrait is as follows: S41. For behavior b, adjust the weights by combining historical data and real-time behavior data: ； In the formula, w t (b) represents the current weight; w t-1 (b) represents the historical weight; w new (b) represents the weight of the latest behavior calculation; γ represents the weight coefficient of historical data, ranging from [0,1]; S42. Dynamically synthesize the interest vector: ； In the formula, represents the updated interest vector; I u Represents the original interest vector; I new represents the interest vector generated by the newly added behavior; α and β represent dynamic weighting coefficients, satisfying α+β=1; S43. Construct a unified feature vector to generate the final user portrait.

5. The method for promoting user personalized content based on big data according to claim 4, characterized in that: The generation process of the final user portrait is as follows: S51. Construct a multi-dimensional user feature vector: ； In the formula, f1, f2, ..., f k Indicates demographic characteristics; S5 2. Add image, text, and voice feature embeddings to enrich the user portrait: ； In the formula, G u Represents the image feature embedding vector; T u Represents the text feature embedding vector; A u Represents the speech feature embedding vector; S53. Reduce the dimension through principal component analysis to improve the calculation efficiency: ； In the formula, represents the user feature vector after dimensionality reduction; W represents the dimensionality reduction mapping matrix; S54. Output end-user portrait .

6. The method for promoting user personalized content based on big data according to claim 1, characterized in that: The generation process of the verification tag is as follows: S61. Randomly select large prime numbers p and q where p ≠ q, and calculate the first-level tag generation factor fg1: fg1 = p × q × (p - 1) × (q - 1); where p > 3, q > 3; S62. Randomly select an integer r and calculate the second-level tag generation factor fg2: ； Among them, p>2 r And q>2 r ; S63. Select the tag parsing code Ca and the tag generation code Cg that satisfy: Cg × Ca = 1 (mod n); where 1 < Ca < fg2, gcd(Ca, fg2) = 1, and gcd() represents the greatest common divisor function; S64, generate verification mark La=H1(M) Cg mod fg1; Among them, M represents the constructed user portrait A binary string; H1() represents a hash function.

7. The method for promoting user personalized content based on big data according to claim 1, characterized in that: The verification process for verifying the rationality of the received user portrait is as follows: S71, calculate the first-level check factor fa1 = (La) Ca mod fg1; where Ca represents the tag parsing code; fg1 represents the first-level tag generation factor; S72. Calculate the second-level verification factor fa2 = H1(M); Among them, M represents the constructed user portrait Binary string; H1() represents the hash function; S73. If fa1=fa2, then the received user profile It is reasonable.

8. The method for promoting user personalized content based on big data according to claim 1, characterized in that: The network structure of the recommendation model that fuses the user portrait and content features includes: S81. The recommendation model integrating user portrait and content features is based on a deep neural network architecture, including: input layer, hidden layer and output layer: ; In the formula, represents the probability of user u's preference for content i predicted by the model; σ represents the activation function; W (L) and b (L) Represent the weight matrix and bias vector of the Lth layer respectively; h (L-1) represents the output of the L-1th layer; S82. The feature fusion layer where user features and content features interact: ； Where W u and W c The projection matrix representing user features and content features; represents a nonlinear activation function; b (1) represents the bias term of the feature fusion layer; h (1) represents the output of the first layer, i.e., the fusion result of user and content features; S83. The hidden layer extracts high-dimensional features: ； In the formula, l represents the number of the hidden layer, l = 2, 3,..., L - 1; S84. Convert the output of the recommendation model into a preference probability through the Sigmoid function: ； In the formula, It represents the probability of user u's preference for content i predicted by the model.

9. The method for promoting user personalized content based on big data according to claim 1, characterized in that: The generation process of the personalized push list is as follows: S91. For each recommended content C i , based on user portraits and the predicted value of the recommended model , calculate the probability of push P push (C i ||u): ； In the formula, λ'' represents the attenuation coefficient; S92, combining the user's click rate, conversion rate and user interaction rate, and combining the objective function L push Select push content: ； Where, L click Indicates click rate loss; L conversion Indicates conversion rate loss; L engagement represents the user interaction rate loss; α', β' and γ' represent the weight coefficients of each objective function respectively; S93. For each user u, calculate a reasonable push frequency f u : ； Where σ' represents the adjustment parameter; C u represents the set of recommended content received by user u; S94: Push probability P based on each content push (C i ||u), user interest prediction value and the properties of the content I(C i ), comprehensively generate personalized push list List u : ； In the formula, I(C i ) represents a content attribute vector, including but not limited to the freshness, category, and tag information of the content; Sort() represents a sorting function.

10. A user personalized content push system based on big data, which is used to execute the user personalized content promotion method based on big data according to any one of claims 1 to 9, characterized in that: It includes: The data collection module is responsible for collecting user behavior and preference data and transmitting it to the data processing module for processing; The data processing module is used to clean the collected user behavior and preference data; The data analysis module is used to accurately model based on user behavior and preference data using clustering analysis and association rule analysis techniques and construct a multi-dimensional user portrait; The personalized recommendation module is used to construct a recommendation model based on the user portrait, content features, and the user's real-time behavior and generate personalized recommendation content; The content push module is used to push personalized content to users through an intelligent push mechanism and can automatically optimize and adjust during the push process.

Images